Open, auditable signatures
Every detection signature is open and auditable–you can read exactly what fires and why–and every one is mapped to MITRE ATLAS techniques and OWASP ASI categories. No black boxes, and no signature counts inflated for a slide. The content is the claim. Two signatures, side by side: one prompt-side, one cloud-side.
id: raxe-pi-001
name: direct_instruction_override
family: prompt_injection
severity: high
confidence: 0.9
stream: S4
mitre_atlas: AML.T0051
owasp_asi: ASI01
patterns:
- "ignore (all |any )?(previous |prior )?instructions"
- "disregard (your |the )?system prompt"
- "you are now (a |an )?[\\w]+ (mode|assistant)"
action: detect
id: raxe-aws-exfil-001
name: Agent Role Reads Production Secret
severity: high
surface: aws-cloudtrail
api_action: secretsmanager:GetSecretValue
identity: agent execution role
finding: data-exfiltration
mitre_atlas: AML.T0025
owasp_asi: ASI04
action: detect
Detections from your own CloudTrail, correlated to agent sessions–not exact attribution. How the AWS lane works →
Where the findings come from
Adversarial ML
Model exploitation, jailbreaks, adversarial inputs.
Agent Security
Tool-use abuse, MCP vulnerabilities, sandbox escapes.
Supply Chain
AI/ML frameworks, registries, pipeline dependencies.
Prompt Injection
Direct, indirect, and web-based injection against agents.
One research pipeline. Five detection signals.
Intelligence feeds every signal in the detection stack–named honestly, one by one.
ATLAS + ASI-Mapped Rules
New techniques from advisories and research streams become open signatures, each tagged to MITRE ATLAS and OWASP ASI.
ML Text Scorer
Research corpora feed the training and evaluation of the text scorer–so verdicts return a threat probability and per-family scores, not just a label.
Structured Tool-Call Risk Analysis
New injection classes found in the wild become argument-level checks: SQL injection, shell injection, and path traversal inside tool-call parameters.
Agent Behaviour Graph
Multi-step attack patterns documented in research inform the sequences the behaviour graph surfaces across an agent session.
Advisory LLM Judge
An optional second opinion on borderline text, tuned against research-derived boundary cases. Evidence, never authority–it never changes the decision.
Put the research to work
Read the research in the open, or see the detection signatures running on your own stack.